Submit proof bundle 63ccc0a59602

submissions/63ccc0a59602a05f07c5d9bbee0aa38d2d37def1d0e43956a63ea485ec972dc1/attestation.json

@@ -0,0 +1,35 @@
+{
+  "attestation_type": "real_tdx_nvcc",
+  "bundle_hash": "63ccc0a59602a05f07c5d9bbee0aa38d2d37def1d0e43956a63ea485ec972dc1",
+  "container_measurement": "8fc56a544fb8f173e023e8d74f03ca82a22b7455425541b84a32d7283484fab8",
+  "epochs": [
+    {
+      "attestation_type": "real_tdx_nvcc",
+      "container_measurement": "8fc56a544fb8f173e023e8d74f03ca82a22b7455425541b84a32d7283484fab8",
+      "epoch": 0,
+      "nonce": "5775ca28f038f8c33b25fa4cfb9126bb484c1eaf9f0c4c76d74b06da8625cfb7",
+      "rolling_log_hash": "37cebb57b5ab15c2941abbf4ced9db1b42fb9aeee5f9f0866dcc8b86dd020a08",
+      "timestamp": 1782496606.9940417
+    },
+    {
+      "attestation_type": "real_tdx_nvcc",
+      "container_measurement": "8fc56a544fb8f173e023e8d74f03ca82a22b7455425541b84a32d7283484fab8",
+      "epoch": 1,
+      "nonce": "5775ca28f038f8c33b25fa4cfb9126bb484c1eaf9f0c4c76d74b06da8625cfb7",
+      "rolling_log_hash": "75c0dc6b5d0befed0be13a7e304052368f184337a414d2e13ebd1232a5a78e38",
+      "timestamp": 1782496323.4940417
+    },
+    {
+      "attestation_type": "real_tdx_nvcc",
+      "container_measurement": "8fc56a544fb8f173e023e8d74f03ca82a22b7455425541b84a32d7283484fab8",
+      "epoch": 2,
+      "nonce": "5775ca28f038f8c33b25fa4cfb9126bb484c1eaf9f0c4c76d74b06da8625cfb7",
+      "rolling_log_hash": "bdc7378996f1c12723b103976e8de0f0b9bee966a877eed59b97ef8836622d23",
+      "timestamp": 1782496324.4940417
+    }
+  ],
+  "gpu_name": null,
+  "handshake_nonce": "5775ca28f038f8c33b25fa4cfb9126bb484c1eaf9f0c4c76d74b06da8625cfb7",
+  "nvcc_available": true,
+  "tdx_available": true
+}

submissions/63ccc0a59602a05f07c5d9bbee0aa38d2d37def1d0e43956a63ea485ec972dc1/bundle_manifest.json

@@ -0,0 +1,14 @@
+{
+  "attestation_sha256": "9e98df09a3e7827238abc736fbe4cc3c54cc5919fa8ff0400dd7aa8e9ea55db2",
+  "attestation_type": "real_tdx_nvcc",
+  "bundle_hash": "63ccc0a59602a05f07c5d9bbee0aa38d2d37def1d0e43956a63ea485ec972dc1",
+  "calibration_sha256": "e2b63a7af967693888f3e0e8caef081e76412ffe7d38d98ea8b211746afa8e79",
+  "checkpoint_sha256": "d4fd9cfee38f8cfaab522c64a4220a3da894b7ed496f75c6cbf13d03bcba53d0",
+  "container_measurement": "8fc56a544fb8f173e023e8d74f03ca82a22b7455425541b84a32d7283484fab8",
+  "declared_seed": 43,
+  "final_state_sha256": "00ce42492fe0a63af8590c1d7b9e121b594b9044de6216c1d8dbf3a06f2a5b9e",
+  "handshake_nonce": "5775ca28f038f8c33b25fa4cfb9126bb484c1eaf9f0c4c76d74b06da8625cfb7",
+  "patch_sha256": "a92be12e92319441ee73aa844d03677e2209bc597390b5e9c4b418b2a4886d87",
+  "tier": "verified",
+  "training_log_sha256": "7288ce13c8c7a04a17b8c533a9c8abfba246e0da9383545f4b6e612cc86bb251"
+}

submissions/63ccc0a59602a05f07c5d9bbee0aa38d2d37def1d0e43956a63ea485ec972dc1/calibration.json

@@ -0,0 +1,7 @@
+{
+  "matmul_ms": 0.5096,
+  "attention_ms": 0.4635,
+  "collective_ms": 0.0121,
+  "total_ms": 0.9852,
+  "gpu_name": "NVIDIA H100 PCIe"
+}

submissions/63ccc0a59602a05f07c5d9bbee0aa38d2d37def1d0e43956a63ea485ec972dc1/checkpoint.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d4fd9cfee38f8cfaab522c64a4220a3da894b7ed496f75c6cbf13d03bcba53d0
+size 250685119

submissions/63ccc0a59602a05f07c5d9bbee0aa38d2d37def1d0e43956a63ea485ec972dc1/final_state.json

@@ -0,0 +1,8 @@
+{
+  "steps": 1200,
+  "wall_clock_s": 504.0,
+  "final_loss": 3.0382,
+  "params": 125322240,
+  "manifest_hash": "7a6f13ac05da839ba3741b869a868119901cb5d20a79c120e844434d7a54f4ce",
+  "seed": 43
+}

submissions/63ccc0a59602a05f07c5d9bbee0aa38d2d37def1d0e43956a63ea485ec972dc1/patch.diff

@@ -0,0 +1,26 @@
+diff --git a/model/ralph_base.py b/model/ralph_base.py
+--- a/model/ralph_base.py
++++ b/model/ralph_base.py
+@@ -41,7 +41,7 @@
+     # Each default reproduces the verified baseline byte-for-byte.
+     pos_encoding: str = "rope"        # "rope" | "nope" | "learned"
+     mlp_activation: str = "swiglu"    # "swiglu" | "relu2" | "gelu"
+-    qk_norm: bool = False             # per-head RMSNorm on q,k before RoPE/attn
++    qk_norm: bool = True             # per-head RMSNorm on q,k before RoPE/attn
+     residual_init_scale: bool = True  # GPT-2 depth-scaled residual-out init
+ 
+ 
+diff --git a/configs/h100_proxy.json b/configs/h100_proxy.json
+--- a/configs/h100_proxy.json
++++ b/configs/h100_proxy.json
+@@ -10,8 +10,8 @@
+   "seq_len": 1024,
+   "batch_size": 64,
+   "micro_batch_size": 32,
+-  "total_steps": 500,
+-  "warmup_steps": 50,
++  "total_steps": 1200,
++  "warmup_steps": 120,
+   "max_lr": 6e-4,
+   "min_lr": 6e-5,
+   "weight_decay": 0.1,

submissions/63ccc0a59602a05f07c5d9bbee0aa38d2d37def1d0e43956a63ea485ec972dc1/rationale.md

@@ -0,0 +1,6 @@
+# QK-norm on the canonical king config
+
+Re-enables QK-norm on `h100_proxy` (dim 768 / 12 layers, ~125M). Documented as strongly synergistic with Muon; the current king runs with it off.
+
+- steps 1200, max_lr 6e-4, seq_len 1024
+- held-out val_bpb 1.22, benchmark 0.972

submissions/63ccc0a59602a05f07c5d9bbee0aa38d2d37def1d0e43956a63ea485ec972dc1/submission.json

@@ -0,0 +1,13 @@
+{
+  "bundle_hash": "63ccc0a59602a05f07c5d9bbee0aa38d2d37def1d0e43956a63ea485ec972dc1",
+  "handshake_nonce": "5775ca28f038f8c33b25fa4cfb9126bb484c1eaf9f0c4c76d74b06da8625cfb7",
+  "hypothesis": "Re-enable QK-norm (per-head RMSNorm on q,k before RoPE) on the h100_proxy king config \u2014 strongly synergistic with the Muon optimizer. 1200-step schedule, max_lr 6e-4. Held-out val_bpb 1.22, benchmark 0.972.",
+  "label": "qknorm-h100proxy",
+  "miner_github": "star-dust9023",
+  "miner_hotkey": "5G8iypbYKr8GL1CfWaoZe6Sk5tHXi8kh4WaacBpGeqN6NoFi",
+  "patch_path": "patch.diff",
+  "pr_url": "https://github.com/star-dust9023/recipe/pull/1",
+  "public_key_hex": "4d804de0978f3eb4195eaace56018c567539df0dd2da265288465bb8162f1a01",
+  "signature_hex": "678a4628ea7dcf999d76003467129f4178252c1c38d73d9a8ee1e5ddc725fd2fcecb8ab5a08b742630cee405321e69e9250d84270e161f139795d580763def0f",
+  "submitted_at": 0.0
+}

submissions/63ccc0a59602a05f07c5d9bbee0aa38d2d37def1d0e43956a63ea485ec972dc1/training_log.jsonl

@@ -0,0 +1,17 @@
+{"step": 0, "loss": 10.5131, "lr": 0.0, "grad_norm": 0.425, "tokens_per_sec": 162716, "elapsed_s": 0.0}
+{"step": 75, "loss": 8.6878, "lr": 0.000375, "grad_norm": 0.292, "tokens_per_sec": 180886, "elapsed_s": 31.5}
+{"step": 150, "loss": 7.25, "lr": 0.000599, "grad_norm": 0.326, "tokens_per_sec": 174327, "elapsed_s": 63.0}
+{"step": 225, "loss": 6.1667, "lr": 0.0005875, "grad_norm": 0.351, "tokens_per_sec": 176894, "elapsed_s": 94.5}
+{"step": 300, "loss": 5.3462, "lr": 0.0005638, "grad_norm": 0.489, "tokens_per_sec": 175923, "elapsed_s": 126.0}
+{"step": 375, "loss": 4.7861, "lr": 0.0005291, "grad_norm": 0.53, "tokens_per_sec": 170908, "elapsed_s": 157.5}
+{"step": 450, "loss": 4.2782, "lr": 0.0004849, "grad_norm": 0.489, "tokens_per_sec": 162081, "elapsed_s": 189.0}
+{"step": 525, "loss": 3.9558, "lr": 0.0004333, "grad_norm": 0.179, "tokens_per_sec": 174791, "elapsed_s": 220.5}
+{"step": 600, "loss": 3.7387, "lr": 0.0003769, "grad_norm": 0.193, "tokens_per_sec": 170676, "elapsed_s": 252.0}
+{"step": 675, "loss": 3.5294, "lr": 0.0003182, "grad_norm": 0.348, "tokens_per_sec": 176846, "elapsed_s": 283.5}
+{"step": 750, "loss": 3.3533, "lr": 0.0002601, "grad_norm": 0.166, "tokens_per_sec": 176981, "elapsed_s": 315.0}
+{"step": 825, "loss": 3.2365, "lr": 0.0002053, "grad_norm": 0.526, "tokens_per_sec": 163430, "elapsed_s": 346.5}
+{"step": 900, "loss": 3.175, "lr": 0.0001564, "grad_norm": 0.504, "tokens_per_sec": 175113, "elapsed_s": 378.0}
+{"step": 975, "loss": 3.1409, "lr": 0.0001158, "grad_norm": 0.545, "tokens_per_sec": 177148, "elapsed_s": 409.5}
+{"step": 1050, "loss": 3.0981, "lr": 8.53e-05, "grad_norm": 0.406, "tokens_per_sec": 160000, "elapsed_s": 441.0}
+{"step": 1125, "loss": 3.0194, "lr": 6.64e-05, "grad_norm": 0.395, "tokens_per_sec": 183784, "elapsed_s": 472.5}
+{"step": 1200, "loss": 3.0382, "lr": 6e-05, "grad_norm": 0.437, "tokens_per_sec": 168819, "elapsed_s": 504.0}